[Paper Review] Towards LLM-based Fact Verification on News Claims with a Hierarchical Step-by-Step Prompting Method

Xuan Zhang, Wei Gao|arXiv (Cornell University)|Sep 30, 2023

Topic Modeling4 citations

TL;DR

This paper proposes a Hierarchical Step-by-Step (HiSS) prompting method that enhances LLM-based fact verification by decomposing complex news claims into subclaims and verifying each via step-by-step reasoning with external search. Evaluated on RAWFC and LIAR datasets, HiSS outperforms fully-supervised state-of-the-art models by 4.95% in macro F1 and generates more accurate, readable, and comprehensive explanations than existing baselines.

ABSTRACT

While large pre-trained language models (LLMs) have shown their impressive capabilities in various NLP tasks, they are still under-explored in the misinformation domain. In this paper, we examine LLMs with in-context learning (ICL) for news claim verification, and find that only with 4-shot demonstration examples, the performance of several prompting methods can be comparable with previous supervised models. To further boost performance, we introduce a Hierarchical Step-by-Step (HiSS) prompting method which directs LLMs to separate a claim into several subclaims and then verify each of them via multiple questions-answering steps progressively. Experiment results on two public misinformation datasets show that HiSS prompting outperforms state-of-the-art fully-supervised approach and strong few-shot ICL-enabled baselines.

Motivation & Objective

• To investigate the effectiveness of in-context learning (ICL) with large language models (LLMs) for news claim verification.
• To address key limitations of vanilla Chain-of-Thought (CoT) prompting—such as thought omission and fact hallucination—in the context of misinformation.
• To design a hierarchical prompting framework that enables fine-grained, step-by-step verification of news claims using LLMs.
• To improve both accuracy and explainability of LLM-based fact verification through structured reasoning and external evidence retrieval.

Proposed method

• The HiSS prompting method consists of two stages: claim decomposition and subclaim verification.
• In claim decomposition, the LLM is prompted to split a complex claim into smaller, logically independent subclaims.
• For each subclaim, the LLM performs step-by-step verification using a chain-of-thought reasoning process.
• External evidence is retrieved via search engines to ground reasoning and reduce hallucination.
• The method uses few-shot in-context learning with four demonstration examples to guide the LLM without fine-tuning.
• Human-annotated explanations are used to evaluate the quality and coherence of generated reasoning chains.

Figure 1: An example of claim verification based on vanilla CoT prompting. The claim (underlined) and CoT (in green) are given as a demonstration. The generated CoT (in italics) leads to an incorrect judgment due to (1) omission of necessary thoughts regarding “nukes”, and (2) fact hallucination abo

Experimental results

Research questions

• RQ1 Can few-shot in-context learning with LLMs achieve performance comparable to fully-supervised models in news claim verification?
• RQ2 Why does vanilla Chain-of-Thought prompting underperform compared to standard prompting in fact verification?
• RQ3 Can hierarchical decomposition of claims improve reasoning completeness and reduce hallucination in LLM-based verification?
• RQ4 To what extent does integrating external search enhance the factual consistency and accuracy of LLM-generated explanations?
• RQ5 How does the explainability of HiSS-generated reasoning compare to human-written explanations and existing automatic baselines?

Key findings

• With only four-shot demonstrations, LLMs using in-context learning achieve performance comparable to strong supervised models on news claim verification.
• Vanilla Chain-of-Thought prompting underperforms due to omission of key claim components and generation of hallucinated facts.
• The HiSS prompting method improves macro F1 by 4.95% on average over state-of-the-art fully-supervised models on two benchmark datasets.
• Human evaluation shows that HiSS-generated explanations are more readable and comprehensive than those from the CofCED baseline, with coverage scores nearly matching human-written explanations.
• HiSS explanations achieve higher readability and non-redundancy than the CofCED model, though they still fall slightly short of gold-standard human explanations.
• The method effectively reduces fact hallucination by grounding reasoning in external search results, though performance depends on the availability of indexed information.

Figure 2: Overview of the proposed HiSS model: Original human inputs are in red background, LLM directly generated text is in white, and answers generated based on search results are in green. We start by providing a few-shot demonstration, followed by appending the claim to be checked (underlined).